DeepRL-Navigation

Banana World

Agent must collect bananas in a large, square room.

Task Type

Episodic

Rewards

+1 for each yellow banana

-1 for each blue banana

Goal

Navigate through the room, collecting as many yellow bananas as possible while avoiding the blue bananas

This environment is considered "solved" when the agent recieves a scoreof 13 points, averaged over 100 episodes

State Space

• State Size

  37 Dimensions

• Space Type

  Continuous

• Representation

  Agent Velocity

  Ray traces to objects around agent's forward direction

Action Space

• Space Size

  4 Dimensions

• Space Type

  Discrete space

• Repesentation

  0	1	2	3
  forward	backward	left	right

Installation

In a new virtual environment:

cd {local path to cloned repo}
cd ./python && python setup.py build && python setup.py install && python setup.py clean && cd -

Running the agent

cd {local path to cloned repo}
python navigation.py {options}

• Available Options

  Option	Description
  --novis	Run without visualization
  --env-dir	Set directory where environment is implemented
  --env-file	Set file for environment
  --n-episodes	Set number of episodes to run agent
  --seed	Set random seed for reproducibility
  --cpu	Use CPU instead of GPU
  --help	Show help message
  --prioritized	Not Yet Implemented Use Prioritized Replay Memory
  --double	Not Yet Implemented Use double deep q network
  --dueling	Not Yet Implemented Use dueling deep q networks